algorithm theoretical basis document (atbd): …

Title: NEON Algorithm Theoretical Basis Document (ATBD): Ecosystem Structure Date: 07/01/2019

NEON Doc. #: NEON.DOC.002387 Author: Tristan Goulden and Victoria Scholl Revision: A

Template_NEON.DOC.000152 Rev D

ALGORITHM THEORETICAL BASIS DOCUMENT (ATBD):

ECOSYSTEM STRUCTURE

PREPARED BY ORGANIZATION DATE

Tristan Goulden AOP 5/29/2019

Victoria Scholl AOP 5/29/2019

APPROVALS ORGANIZATION APPROVAL DATE

David Barlow SYS 06/11/2019

RELEASED BY ORGANIZATION RELEASE DATE

Anne Balsley CM 07/01/2019

See configuration management system for approval history.

The National Ecological Observatory Network is a project solely funded by the National Science Foundation and managed under cooperative agreement by Battelle.

Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the

National Science Foundation.



Template_NEON.DOC.000152 Rev D

Change Record

REVISION DATE ECO # DESCRIPTION OF CHANGE

A 07/01/2019 ECO-06170 Initial Release


NEON Doc. #: NEON.DOC.002387 Author: Tristan Goulden Revision: A

Page 1 of 28

TABLE OF CONTENTS

1 DESCRIPTION ..................................................................................................................................3

1.1 Purpose ............................................................................................................................................... 3

1.2 Scope ................................................................................................................................................... 3

2 RELATED DOCUMENTS, ACRONYMS AND VARIABLE NOMENCLATURE ..............................................4

2.1 Applicable Documents ....................................................................................................................... 4

2.2 Reference Documents ........................................................................................................................ 4

2.3 Acronyms ............................................................................................................................................ 4

3 DATA PRODUCT DESCRIPTION .........................................................................................................5

3.1 Variables Reported ............................................................................................................................. 5

3.2 Input Dependencies ........................................................................................................................... 5

3.3 Product Instances ............................................................................................................................... 5

3.4 Temporal Resolution and Extent ....................................................................................................... 5

3.5 Spatial Resolution and Extent ............................................................................................................ 6

4 SCIENTIFIC CONTEXT........................................................................................................................6

4.1 Theory of Measurement .................................................................................................................... 7

4.2 Theory of Algorithm ........................................................................................................................... 8

5 ALGORITHM IMPLEMENTATION .................................................................................................... 13

6 UNCERTAINTY ............................................................................................................................... 15

6.1 Observed CHM uncertainty.............................................................................................................. 18

7 VALIDATION AND VERIFICATION .................................................................................................... 22

8 FUTURE PLANS AND MODIFICATIONS ............................................................................................ 24

9 BIBLIOGRAPHY .............................................................................................................................. 24

LIST OF TABLES AND FIGURES

Table 1 - Data products generated by algorithms described within this ATBD ................................................... 5

Figure 1 - CHM and RGB image of same area ..................................................................................................... 8



Page 2 of 28

Figure 2 - Left: CHM without algorithm applied to remove data pits, identifiable by black pixels. Right:

CHM with algorithm applied to remove data pits............................................................................................. 11

Figure 3 - Top - Khosravipour, Skidmore, Isenburg, Wang, and Hussin (2014) algorithm for creation of a pit-

free CHM. Bottom - NEON adaptation of the Khosravipour, Skidmore, Isenburg, Wang, and Hussin (2014)

algorithm for creation of a pit-free CHM to allow for a dynamic height ceiling and selectable height intervals

.......................................................................................................................................................................... 12

Figure 4 - Flowchart summarizing CHM creation ............................................................................................. 14

Figure 5 - Left: coverage of BRDF flight at SOAP. Right: coverage of BRDF flight at SJER .................................. 18

Figure 6 - Left: sample of uncertainty of CHM at SOAP. Right: sample of uncertainty of CHM at SJER ............ 20

Figure 7 - Top: histogram of cell by cell standard deviation of the CHM at SOAP. Bottom: histogram of cell

by cell standard deviation of the CHM at SJER .................................................................................................. 21

Figure 8 - Linear regression of ground measured canopy heights vs. the LiDAR derived CHM at SJER ............ 23

Figure 9 - Vertical residuals of ground measured tree height vs. LiDAR derived CHM linear regression model

at SJER ............................................................................................................................................................... 24



Page 3 of 28

1 DESCRIPTION

1.1 Purpose

This document details the algorithms used for creating the NEON Level 3 ecosystem structure data product

(NEON.DOM.SITE.DP3.30015.001) from Level 1 data, and ancillary data (such as calibration data), obtained via

instrumental measurements made by the Light Detection and Ranging (LiDAR) sensor on the Airborne

Observation Platform (AOP). It includes a detailed discussion of measurement theory and implementation,

appropriate theoretical background, data product provenance, quality assurance and control methods used,

approximations and/or assumptions made, and a detailed exposition of uncertainty resulting in a cumulative

reported uncertainty for this product.

1.2 Scope

This document describes the theoretical background and entire algorithmic process for creating

NEON.DOM.SITE.DP3.30015.001 from input data. It does not provide computational implementation details,

except for cases where these stem directly from algorithmic choices explained here.



Page 4 of 28

2 RELATED DOCUMENTS, ACRONYMS AND VARIABLE NOMENCLATURE

2.1 Applicable Documents

AD[01] NEON.DOC.000001 NEON Observatory Design (NOD) Requirements

AD[02] NEON.DOC.002652 NEON Level 1, Level 2 and Level 3 Data Products Catalog

AD[03] NEON.DOC.002293 NEON Discrete LiDAR datum reconciliation report

AD[04] NEON.DOC.002649 NEON configured site list

2.2 Reference Documents

RD[01] NEON.DOC.000008 NEON Acronym List

RD[02] NEON.DOC.000243 NEON Glossary of Terms

RD[03] NEON.DOC.002390 NEON elevation Algorithm (DTM and DSM) Theoretical Basis

document

RD[04] NEON.DOC.001984 AOP flight plan boundaries design

RD[05] NEON.DOC.005011 NEON Coordinate Systems Specification

RD[06] NEON.DOC.001292 NEON L0-to-L1 discrete return lidar algorithm theoretical basis

RD[07] NEON.DOC.002890 NEON AOP Level 0 quality checks

RD[08] NEON.DOC.003791 NEON Elevation (Slope and Aspect) Algorithm Theoretical Basis

Document

RD[09] NEON.DOC.00987 TOS Protocol and Procedure: Measurement of Vegetation Structure

RD[10] NEON.DOC.002186 AOP determination peak greenness plan

2.3 Acronyms

Acronym Explanation

DTM Digital Terrain model

DSM Digital Surface model

DEM Digital Elevation Model

ITRF00 International Terrestrial Reference Frame 2000

UTM Universal Transverse Mercator

TIFF Tagged Image File Format

AOP Airborne Observation Platform

FBO Fixed Base Operator

PPM pulses per square meter

CHM Canopy Height Model

HDF Hierarchical data format

AGC Above Ground Carbon

AGB Above Ground Biomass

SOAP Soaproot Saddle



Page 5 of 28

SJER San Joaquin Experimental Range

ASPRS American Society of Photogrammetry and Remote Sensing

3 DATA PRODUCT DESCRIPTION

3.1 Variables Reported

The product supplied through NEON.DOM.SIT.DP3.30025.001 includes a canopy height model (CHM) in raster data format. The CHM is derived from the lidar point cloud, a level 1 product (see RD[06]). Rasters for the CHM are reported with horizontal reference to the ITRF00 datum, projected to the Universal Transverse Mercator (UTM) Cartesian coordinate system in accordance with RD[06]. The CHM is a normalized by ground height, indicating all ground elevations are set to zero and canopy height is the height of vegetation above the ground in meters. The CHM rasters are divided into a set of 1 km by 1 km tiles, which have corners spatially referenced to an even kilometer. The product is stored in a GeoTIFF format in accordance with the GeoTIFF specification (Ritter et al., 2000).

3.2 Input Dependencies

The creation of the canopy height model rasters requires tiled LiDAR point clouds in LAS format as input.

Procedures for creating the L1 point cloud, including tiling, can be found in RD[06] and procedures for pre-

processing of the point cloud in RD[03].

3.3 Product Instances

The NEON data products produced directly from these algorithms are summarized in Table 1.

Table 1 - Data products generated by algorithms described within this ATBD

Data product identification Data product name

NEON.DOM.SITE.DP3.30015.001 Canopy height model

3.4 Temporal Resolution and Extent

The canopy height model product will include data collected during acquisition of a single core, re-locatable or

aquatic site by the AOP. Depending on external variables such as weather, transit time to the site FBO, and

total area of the priority 1 flight box (see RD[04]), the temporal resolution of a single acquisition of L0 LiDAR

information could range from a single flight (4 hrs.) to several flights acquired over multiple days. Generally,

given AOP intends to acquire data from each site during a period of peak greenness (RD[12]) annually, the

total potential time to acquire a site will have a limit which defines the largest temporal resolution for a single

acquisition. Details defining the total amount of potential time dedicated to a single site acquisition are

provided in RD[04]. As the NEON AOP payload is scheduled to repeat each NEON site on an annual basis, the

temporal resolution of multiple acquisitions will be one year.



Page 6 of 28

3.5 Spatial Resolution and Extent

The canopy height models will be created at 1 m spatial resolution and be spatially coincident with the DTM,

DSM (RD[08]), slope and aspect maps (RD[10]). The planned spatial extent of the canopy height model will

relate to the definition of the AOP flight box (RD[04]) for each individual site (RD[04]). It is intended that a

minimum of 80% of the priority 1 flight box and 95% of the tower airshed will be acquired each year

(RD[07]). As discussed in Section 3.4, the actual acquired area could vary depending on external conditions

encountered during the flight. Ultimately, the flight schedule as defined in RD[04] shall supersede the percent

coverage requirement. Therefore, the actual acquired spatial extent may vary annually.

4 SCIENTIFIC CONTEXT

Recent global initiatives aimed at reducing greenhouse gas emissions have generated interest in

quantification and monitoring of forest inventories. Forests store and sequester a considerable proportion of

the terrestrial global carbon budget; therefore, policy actions which incentivize new growth can reduce

current available atmospheric carbon introduced by anthropogenic greenhouse gas emissions. Conversely,

carbon can also be released into the atmosphere thorough activities such as deforestation, forest degradation

or burning which can undermine the effectiveness of policy aimed at reducing the quantity of atmospheric

carbon. Climate change legislation as early as the Kyoto Protocol recognized the utility of accounting for

carbon sinks associated with vegetation growth as a viable method to quantify an offset in carbon emissions

(Rosenqvist et al., 2003), and to monitor the global carbon budget. Contemporary international efforts aimed

at carbon accounting such as REDD+ (Reducing Emissions from Deforestation and forest Degradation) target

monitoring carbon emissions from land-use change due to deforestation.

Remote sensing has been identified as one of the primary tools which will enable global monitoring of

forest stocks to support carbon monitoring through changes in land-use (i.e. deforestation) (Rosenqvist et al.,

2003; Goetz & Dubayah, 2011). In particular, LiDAR remote sensing has gained recent interest as a tool for

predicting above ground biomass (AGB), a primary indicator of carbon storage potential, and a proxy to estimate

carbon contributions of ground litter and below-ground biomass (Goetz & Dubayah, 2011). LiDAR sensors have

shown higher success over optical sensors in areas of high density biomass and LAI, where optical sensors tend

to lose accuracy and sensitivity (Lefsky et al., 2002; Turner et al., 1999; Carlson & Ripley, 1997), and do not

directly measure three dimensional forest structure. Forest canopy metrics are directly measurable with

LiDAR sensors bebecause laser pulses will be reflected from the uppermost canopy layers and remaining

energy will penetrate to, and reflect from, under-story and the ground surface. The near simultaneous direct

measurement of ground and canopy elevation allows the canopy height to be estimated through

straightforward differencing. A CHM is then generated by creating a continuous surface of canopy height

estimates across the entire spatial domain of the LiDAR survey.

Goetz and Dubayah (2011) state that canopy height is one of the most desirable and common metrics used

for ecosystem studies. To support carbon accounting initiatives, above ground carbon (AGC) is typically



Page 7 of 28

estimated from estimates of AGB, which can be determined with the support of a CHM. Prior to the

availability of LiDAR, AGB was traditionally inferred from allometric relationships derived from field sampling

of individual trees in spatially distributed plots across the forest of interest. Field sampling of forest plots is

expensive and time consuming, and can result in a lack of adequate sampling in distinct forest biomes, which

will bias the extrapolation of results across a diverse forest ecosystem. The combination of strategically placed

field plots along with the increased spatial coverage of LiDAR aerial surveys is the subject of on-going research

for estimating AGB on the scale of an entire forest system (e.g. Asner et al. (2010)). It has been shown that the

combination of both the highly detailed field plot data and large spatial coverage of LiDAR data allows statistical

regression or machine-learning techniques to accurately spatially extrapolate information outside of forest

plots. Several studies have demonstrated the use of LiDAR-derived CHM metrics for prediction of AGB. For

example, Asner et al. (2010) utilized canopy height as one of several canopy metrics metrics derived from LiDAR

to estimate the AGB over a 4.3 million ha area of the Peruvian Amazon, demonstrating the efficiency in

leveraging LiDAR in combination with field plots for large area mapping of AGC. Asner et al. (2011) also utilized

a similar methodology to map AGB on the million Ha Island of Hawaii. Lefsky et al. (2002) analyzed three distinct

temperate forest biomes (two temperate, one boreal) and found a statistically significant relationship (p <

0.001) between the mean canopy height and field measured above ground biomass in all sites. The

demonstrated utility of LiDAR derived CHMs for assessing AGB indicate the utility in supporting the scientific

objectives of the NEON observatory.

4.1 Theory of Measurement

The CHM is derived directly from the LiDAR point cloud. RD[06] describes in detail the theory of measurement

and observation of the LiDAR point cloud. To summarize, the LiDAR point cloud is produced from LiDAR return

signals from both surface features and the true-ground as LiDAR pulses will be reflected from the uppermost

layers of the canopy, as well as the underlying ground surface. To produce the CHM, the point cloud is

separated into classes representing the ground and vegetation returns. The ground classified points allow

calculation of a height normalized point cloud to provide a relative estimate of vegetation elevation. A surface

is then generated using the height normalized vegetation points to produce the CHM (Figure 1).



Page 8 of 28

Figure 1 - CHM and RGB image of same area

4.2 Theory of Algorithm

Fundamentally, a CHM can be considered a surface of normalized vegetation heights. In the simplest sense,

the CHM can be calculated as a difference between the DSM and the DTM (see RD[08] for background on

DSM and DTM creation):

𝐶𝐻𝑀 = 𝐷𝑆𝑀 − 𝐷𝑇𝑀 (1)

Directly subtracting the DTM from the DSM to determine a CHM can introduce artifacts into the CHM known

as data pits. Data pits manifest as abnormally low elevation pixels within a tree crown and underestimate

the true canopy height, and are a commonly observed problem in CHMs derived from LiDAR (Khosravipour

et al., 2014; Ben-Arie et al., 2009; Liu & Dong, 2014). As the CHM results are dependent on non-linear

interactions between LiDAR acquisition parameters and post-processing routines, the causes of data pits has

not been conclusively proven within existing literature (see Khosravipour et al. (2014) for a summary of

documented causes). Regardless of the cause, it is well known that airborne LiDAR inherently underestimates

canopy height (see Section 6), and the presence of data pits will cause further underestimation of height

metrics derived from a CHM. Bias in height metrics derived from the CHM will lead to errors in estimates of

AGB, carbon sequestration potential (Ben-Arie et al., 2009), or structural forest metrics such as basal area

and stand volume (Shamsoddini et al., 2013; Zhao et al., 2013). Attempts at morphological filtering such as

mean, median or Gaussian have been successful in removing pits; however, filtering techniques will also

undesirably modify the remaining pixels, and potentially reduce CHM accuracy (Ben-Arie et al., 2009;

Khosravipour et al., 2014). Therefore, an algorithm which provides a methodology for removing data pits

while maintaining the fidelity of the original CHM is desirable.



Page 9 of 28

The algorithm implemented by NEON for determining a CHM is detailed in Khosravipour et al. (2014, 2013)

and was designed with the specific intention of removing ’data pits’ in CHMs. The algorithm generates a

standard CHM (sCHM) by creating a surface through triangulation of all normalized first returns, as well as a

series of partial CHMs (pCHMs) by triangulating all first returns above a set of increasing elevation thresholds.

Background into the triangulation algorithm can be found in RD[08]. The maximum value for any given x-y

location (pixel) in the pCHMs and sCHM is used to produce the final pit-free CHM (Figure 3). The underlying

assumption is that as the elevation threshold is increased, only returns from a unique tree crown are included

in pCHM generation (Khosravipour et al., 2014). If data pits exist in the standard CHM, the interpolation

inherent in the TIN algorithm will result in filled pits in the pCHMs. The maximum length allowed for any

given triangular edge in the TIN algorithm is a critical parameter, as it controls the size of data pits that will

be filled, but also allows legitimate gaps in the canopy or empty space between adjacent tree crowns to be

filled. Khosravipour et al. (2014) suggest that the maximum length of triangular edges, referred to as the

rasterization threshold, is larger than the average point spacing between LiDAR observations, yet below the

space between individual trees. For the typical point spacing of NEON LiDAR acquisitions it was determined

that a rasterization threshold of 3 m was reasonable. Values tested below a rasterization threshold of 3 m

resulted in a large number of remaining data-pits. As the optimum value for the rasterization threshold is

likely related to the vegetation structure, additional testing is on-going to develop a method to determine a

dynamic rasterization threshold based on both the LiDAR acquisition parameters and the vegetation

structure.

In the example case study presented in Khosravipour et al. (2014), a static maximum height ceiling threshold

of 20 m was implemented when calculating the pCHMs. The 20 m limit was selected because only 5% of trees

within their study area were taller than 20 m. Due to the differences in vegetation species and growing

conditions across NEON sites, considerable variation exists among the maximum tree height. To apply the

Khosravipour et al. (2014) algorithm to the diverse cross-section of vegetation conditions at NEON sites, the

maximum height ceiling threshold (H) is dynamically selected from the sCHM (Figure 3). Several values of

potential candidates for the ceiling height threshold from the sCHM were tested, 1) the maximum height

value, 2) the 99th percentile of height and 3) the 95th percentile of height. It was determined that the

maximum height within the sCHM is often associated with a non-vegetative surface feature (despite data

filtering in the point cloud pre-processing), such as a noise point above the canopy, or a manufactured

structure such as a communications tower. Therefore, the maximum height from the sCHM is not an efficient

choice for the ceiling height threshold because several levels of pCHMs will be created unnecessarily. Testing

of the 99th and 95th percentile of heights revealed that the 99th percentile was typically associated with

legitimate vegetation, and was often several height bins above the 95th percentile. Therefore, the 99th

percentile of height from the sCHM is currently used to determine the height ceiling threshold.

The implementation of the algorithm in Khosravipour et al. (2014) also had statically selected height bins of

2, 5, 10, 15 and 20 m above which each pCHM was created. The NEON algorithm implementation also uses

2 m as the lowest level, but allows a variable increment value (i) for each subsequent increase to the height

bin. Testing has shown that 5m was a reasonable value for i, however ongoing tests will be conducted to



Page 10 of 28

determine if i can be optimized for different forest environments. The thresholds on the height bins will

proceed in the same pattern as those in Khosravipour et al. (2014), but increase up to the first even interval

of above H (see Figure 3) according to the following equation:

ℎ = [2, 𝑖, 2𝑖, … , 𝑛𝑖] (2)

where h is the vector of height bin thresholds used for each pCHM, n is the total number of required bins

determined as

𝑛 = [𝐻𝑖⁄ ] (3)

4.2.1 Pre-Processing

Prior to creation of the CHM, each point within the LiDAR point cloud must be assigned to a class (Figure 4).

Processing steps implemented up to and including the point classification can be found in RD[06] and RD[08].

The LAStools soft- ware suite (http://www.cs.unc.edu/~isenburg/lastools/) is used for determining the point

classification, and separates points into five potential classes including: 1) unclassified, 2) noise, 3) ground, 4)

high vegetation, and 5) building. The classes of each point are saved as an integer identifier according to

ASPRS (2009) and provided in tiled point clouds in compressed LAS (LAZ) format. As described in RD[08],

internal testing revealed that a large number of points classified as ’unclassified’ belong to the vegetation

class. Therefore, all point data labeled as ’high vegetation’ and ’unclassified’ are used for creating the sCHM

and pCHMs. Only points associated with the ’ground’ class are used for normalization of the elevations in

the ’high vegetation’ and ’unclassified’ points.

Prior to creation of the normalized vegetation heights, the data points are thinned to include only the

maximum return height within a pre-defined area (Figure 4). Thinning the points follows the

recommendation given in Isenburg (2014). LiDAR generated CHMs are known to systematically under-

estimate the true tree canopy (see Section 6), therefore thinning and retaining only the maximum height

reduces the opportunity for CHM underestimation. Point thinning is implemented by superimposing a grid

with 0.5 m spatial resolution over the point cloud, and selecting only the maximum height within each grid

cell. As NEON LiDAR acquisition parameters generally result in ~4 pts /m2 (~1 pt/ 0.5 m2), the thinning at

this resolution will not remove a large percentage of the acquired points. Overlap regions of adjacent flight

lines where the point density is higher than 4 pts /m2 will be more severely altered from the original point

cloud due to the thinning procedure.

http://www.cs.unc.edu/~isenburg/lastools/



Page 11 of 28

Figure 2 - Left: CHM without algorithm applied to remove data pits, identifiable by black pixels. Right: CHM with algorithm applied to remove data pits



Page 12 of 28

Figure 3 - Top - Khosravipour, Skidmore, Isenburg, Wang, and Hussin (2014) algorithm for creation of a pit-free CHM. Bottom - NEON adaptation of the Khosravipour, Skidmore, Isenburg, Wang, and Hussin (2014) algorithm for creation of a pit-free CHM to allow for a dynamic height ceiling and selectable height intervals



Page 13 of 24

5 ALGORITHM IMPLEMENTATION

The processing of the point cloud into the CHM product is achieved through the steps outlined in this

section (Figure 4). The algorithm for CHM creation is outlined through multiple interconnected LAStools

(http://www.cs.unc.edu/~isenburg/lastools/) and in-house Matlab

(http://www.mathworks.com/products/matlab/) functions which automate the algorithm in a cygwin

(https://www.cygwin.com/) environment. The process is dependent on the existence of a point cloud

which has been tiled, noise filtered, and classified into ground and non-ground points (see RD[08]).

Step 1:

Classify the LAS data into ground and non-ground points (see Section 4.2.1).

Input:

1. Ground / non-ground classified point cloud, 1 km by 1 km tiles (with buffer) in LAZ

format (see RD[08]),

2. Flag which identifies the search mode (set to extra fine),

3. Flag which indicates the size of a search area for identification of original ground

points (set to 10),

4. Planer flag (set to 0.1 m).

Output: Tiled LAS files with ground points classified according to ASPRS classification scheme

(ASPRS, 2009).

Functions used: lasground (lastools).

Step 2:

Determine point cloud heights normalized to the ground surface.

Input: Classified point cloud, 1 km by 1 km tiles (with buffer) in LAZ format.

Output: Height normalized point cloud, 1 km by 1 km tiles (with buffer) in LAZ format.

Functions used: lasheight (lastools).

Step 3:

Thin the point cloud (see Section 4.2.1)

Input:

1. Height normalized point cloud, 1 km by 1 km tiles (with buffer) in LAZ format,

2. Step size flag to indicate grid cell size (set to 0.5 m),

3. ’Highest’ flag to indicate to keep the point with maximum elevation in each grid cell.

Output: Height normalized and thinned point cloud, 1 km by 1 km tiles (with buffer) in LAZ

format.

Functions used: lasthin (lastools).

Step 4:

Calculate height thresholds for each pCHM

Input: CHM rasters, 1 km by 1 km tiles in gtif format.

Output: Height thresholds, ascii file.

Functions used: extract_height_from_chm.m (Matlab)



http://www.mathworks.com/products/matlab/

https://www.cygwin.com/



Page 14 of 24

Figure 4 - Flowchart summarizing CHM creation



Page 15 of 24

Step 5:

Calculate pCHMs for each height interval. This step is repeated for each elevation threshold in

the ASCII output from Step 4

Input:

1. CHM rasters, 1 km by 1 km tiles in gtif format

2. Step size flag to indicate spatial resolution (set to 1 m),

3. Elevation threshold calculated from Step 4,

4. Elevation flag to indicate the height normalized elevation is the variable used to create

the TIN,

5. Kill triangles flag to indicate the maximum size of a triangular edge to be maintained

(set to 3 m),

6. Keep class flag to indicate which point classifications to use (unclassified, ground and

vegetation points).

Output: pCHM rasters for each height threshold, 1 km by 1 km tiles, gtiff format.

Functions used: las2dem (lastools).

Step 6:

Combine pCHMs by choosing maximum height for each cell

Input: pCHM rasters, 1 km by 1 km tiles in gtif format from Step 5.

Output: Pit-free CHM, 1 km by 1 km tiles in gtiff format.

Functions used: combine_CHM_gtif.m (Matlab)

6 UNCERTAINTY

It has been generally observed that canopy height is underestimated by airborne laser scanning

observations (see Gaveau and Hill (2003), Maltamo et al., (2004), Suárez et al., (2005), Chasmer et al.,

(2006)). Hyyppä et al. (2009) identify several sources of uncertainty in the creation of canopy height

models which can lead to a systematic underestimation in canopy height; however these sources can also

introduce random errors which can cause variable levels of uncertainty:

1. density / coverage of the the LiDAR points (affected by flying altitude, flying speed, laser scan rate,

pulse repetition frequency (PRF)),

2. laser beam parameters such as beam divergence,

3. algorithm used to classify ground points and create the DTM,

4. sensitivity of the return signal detection hardware processing techniques,

5. the structure of the vegetation.



Page 16 of 24

Analysis of source 1, the density / coverage of LiDAR points, has found that a decrease in point density

results in a systematic reduction in canopy height. The reduction in canopy height due to a decrease in

point density is primarily attributed to a diminished opportunity for the energy from any given pulse to

intercept with an individual tree crown apex (Lefsky et al., 2002). Following this logic, increasing the beam

divergence of the laser pulse will enlarge the sampled area and increase the likelihood the pulse will

intercept maximum points on the canopy. NEON selects the widest available beam divergence setting on

the Optech Gemini system to increase the spatial coverage of the intercepted are, ensure eye-safe

conditions, and increase the likelihood of making contact with tree crown apexes. However, in well-

controlled experiment that varied flight acquisition parameters, Hopkinson, 2007 determined that it was

the peak pulse power concentration (driven by flight altitude, beam divergence, and PRF) that was the

dominant control on systematic variation in laser pulse return distribution within vegetation (related to

source 2 and source 4). A pulse with lower peak power concentration will tend to penetrate deeper into

the foliage before sufficient energy is returned to overcome a minimum threshold required in the receiving

hardware to trigger a return signal. Similar results were also observed by Andersen et al. (2006), who

determined the wide beam divergence setting resulted in a larger underestimation of tree height. Given

that the NEON LiDAR system is operated at one of the higher PRF rates for the Optech Gemini instrument

(100 kHz) and in wide beam divergence mode, the peak power concentration of individual pulses will be

low compared to alternative available LiDAR acquisition parameters and the canopy height will may be

more severely underestimated.

Other studies have noted that some LiDAR acquisition parameters play a minor or negligible role in the

determination of canopy height. For example, Næsset (2004) noted very little difference on mean tree

height with changing flying altitudes and beam divergence. However, the study was conducted with only

300 m of height difference (540 vs 840 m) and with a system only capable of a 10 kHz PRF (ALTM 1210).

Therefore, the differences in peak power concentration due to changes in altitude at a 10 kHz PRF may

have been insufficient to create a notable change in pulse penetration. Næsset (2009) also noted a slight

increase in the canopy height with increasing platform altitude (1100 m vs 2000 m) using the same model

LiDAR sensor as Hopkinson, 2007 (ALTM 3100), which contradicted results from Hopkinson, 2007. Similar

conclusions were also found in Goodwin, Coops, and Culvenor (2006), who found that canopy height

values increased slightly as the platform increased in elevation (1000 m vs 3000 m). In response to the

seemingly conflicting results found in the literature, Næsset (2009) identifies the importance of the

coupling of effects of the different LiDAR parameters, i.e changes in pulse density and beam envelope size

(therefore peak power concentration) change with altitude and PRF. As a result, it is difficult to design a

controlled experiment which conclusively isolates each of the confounding factors. Without knowledge

of the return pulse triggering algorithm and details of hardware in the receiving electronic circuits

(typically proprietary information), it is difficult to draw sound conclusions from purely empirical

experimentation. Including the potential issues associated with source 5, the structure of the vegetation,

additional research is required to fully describe how uncertainty due to LiDAR acquisition parameters will

contribute to uncertainty in canopy height.



Page 17 of 24

The algorithm used to create the DTM (source 3) affects the CHM because the vegetation heights are

normalized to the DTM surface. Large errors can occur in the DTM if pulses are unable to penetrate to the

ground and large areas must be interpolated in order to create a continuous DTM surface. This is

particularly problematic in areas with a dense vegetation where it is unlikely that sufficient pulse energy

will transmit through the canopy and allow a ground return. Topographic features such as valleys, which

commonly exist in riparian areas with dense vegetation, could be interpolated and filled, creating a

systematically raised ground surface and underestimation of the canopy height. Although higher PRFs

result in lower peak pulse power concentrations, Hopkinson (2007) and Chasmer et al., (2006) found that

higher PRFs result in a denser sample of the ground surface at their study sites, reducing the opportunity

for uncertainty to be introduced through interpolation to a DTM. Returns from low vegetation may be

erroneously classified as ground points (Lefsky et al., 2002). This will lead to the canopy height being

determined between the upper levels of the understory and the canopy top and introduce a systematic

underestimation. This effect will be particularly problematic in understory vegetation which is below 2

m in height as this is the minimum detection threshold distance between targets. Therefore, if a return

signal from understory which is below 2 m in height is detected, it is not possible for the sensor to also

detect the ground surface and the low vegetation will appear as the ground surface.

Within the literature, different levels of uncertainty / error have also been reported across various tree

species (source 5). Due to the numerous uncontrolled variables across experiments including flight

parameters and LiDAR systems, it is difficult to synthesize a universal conclusion about the effect of tree

species on canopy height model uncertainty. It has been suggested that both the structure of the canopy

and the reflectivity of the foliage at the given laser wavelength play a role in the returned energy

signature (Lefsky et al., 2002; Hopkinson, 2007), and in canopy height uncertainty. For example, Yu et al.,

(2004) found that the birch trees were less sensitive to changes in altitude than coniferous trees. In a study

of a mixed forest with heterogeneous stands, Maltamo et al., 2004 found different levels of

underestimation of canopy height by tree species. Plot data was available for Norway spruce, Scots pine

and birch, and mean height was underestimated by 0.80, 1.20, and 0.91 meters for each species

respectively. In a study of varying LiDAR acquisition parameters, Hopkinson (2007) noted systematically

varied levels of foliage pulse penetration between vegetation using an Optech ALTM 3100 system at a

high PRF (100 kHz). He concluding that the vegetation type (birch, mixed pine, mixed spruce) exerted

some control over the level of pulse penetration into the canopy. Given the NEON project is designed to

collect data across a range of representative tree species within the continental U.S., it should not be

expected that the level of uncertainty in the CHM will be constant despite consistent collection protocols

with identical acquisition parameters. Additional ground data is required to begin assessing variations in

uncertainty across the various forest types being acquired in the NEON project to understand the role of

different tree species.



Page 18 of 24

6.1 Observed CHM uncertainty

An opportunity to provide an estimate of the uncertainty in CHMs derived from the flight acquisition

parameters and processing techniques utilized at NEON in an applied context was provided through

engineering surveys acquired for SJER (San Joaquin Experimental Range) and SOAP (Soaproot Saddle),

both in NEON’s Domain 17 (D17, http://www. neonscience.org/science-design/spatiotemporal-design).

Specifically, the engineering surveys at these sites included acquisitions intended to analyze BRDF (bi-

directional reflectance effects) of the data acquired by the imaging spectrometer. The BRDF flight plan

collects a portion of the site multiple times with different view geometries through several sets of offset

parallel lines acquired in varied directions. Information about the specifics of the BRDF flight plan can be

found in RD[09]. The repeated coverage allows multiple CHMs to be created on the same spatial grid, one

from each overlapping flight line. The uncertainty in the CHM can then be determined by analyzing the

variability of each individual overlapping grid cell. The area with the highest amount of overlap in each of

the BRDF flights resulted in 18 and 20 samples for the SJER and SOAP BRDF flights respectively (Figure 5).

CHMs were created from each individual flight line using the NEON algorithm (Figure 4). A rectangular clip

area which contained the majority of overlap between the lines was created and used to clip each flight

line. It was noted in processing that interpolation was occurring along the irregularly shaped line edges.

Interpolated areas along the line edges will bias estimates of the CHM variability because these areas

will appear to be ground points, but are truly areas with no legitimate data. To eliminate this issue, the

edge of each flight line was decreased by 5 m, eliminating all interpolated areas. The standard deviation

was then determined for each cell as an estimate of the uncertainty (Figure 6).

Figure 5 - Left: coverage of BRDF flight at SOAP. Right: coverage of BRDF flight at SJER

http://www.neonscience.org/science-design/spatiotemporal-design

http://www.neonscience.org/science-design/spatiotemporal-design



Page 19 of 24

At SOAP, 95% of cells resulted in a standard deviation which was below ±10.5 m, and at SJER 95% of cells

were below ±3.2 m (Figure 7). As the flight acquisition parameters were constant between the two BRDF

flights, differences in uncertainty are likely due to the differences in vegetation structure / density and

reflectance at each site. SJER is dominated by blue oaks, while SOAP is dominated by ponderosa pine and

oak. The canopy height is generally higher at SOAP with 99% of the CHM cells below ~43 m, while at SJER

99% of the CHM cells are below ~18 m. The highest levels of uncertainty occur in SJER at the edge of tree

crowns, while the uncertainty is generally lower within tree crowns (Figure 7). Uncertainty at crown edges

could be a result of a mis-registration in the individual CHM rasters. However, the planimetric accuracy of

the overlapping lines were checked and did not reveal identifiable horizontal shifts. The uncertainty at

the crown edges is likely a consequence of the wide beam divergence mode, which can allow a return to

be triggered from anywhere within the beam envelope (Goulden & Hopkinson, 2010). Some of the BRDF

lines would result in beams that initiated first contact at the edge of the canopy with the outermost

portion of the beam, systematically shifting the crown outward to the location of the beam center. If on

a subsequent line the beam made first contact with the tree crown with the beam center within the tree

crown, the crown would not be extended outward, and the adjacent cell would be considered a ground

pixel. Therefore, edge of canopy pixels could result in ground cells in one line, while being canopy cells in

subsequent lines, resulting in a large standard deviation at crown edges. Cells on the interior of the tree

will rarely be considered ground cells due to the pit-filling algorithm, which eliminates this issue. As the

canopy is relatively open at SJER compared to SOAP, the edge effect is regularly observed and the majority

of cells with a high standard deviation (> 1.35 m) were observed at crown edges (Figure 7). At SOAP, the

uncertainty appears to be highest on trees which are isolated within stands of several shorter trees. The

high uncertainty in these regions of SOAP could be related to the different view geometries of the BRDF

lines. If the scanner makes contact with the tall isolated tree at the higher angle of incidence, it is possible

that the pulses intercept lower portions of the canopy and trigger a first return well below the apex, a

phenomenon identified in Næsset et al. (2004). This appears plausible given the shape of Ponderosa pine,

which have narrow conical crowns. This small experiment displays the levels and patterns of spatial

uncertainty that can exist between two sites given the NEON acquisition parameters and processing

techniques. Additional research is required to further characterize uncertainty at other NEON sites.



Page 20 of 24

Figure 6 - Left: sample of uncertainty of CHM at SOAP. Right: sample of uncertainty of CHM at SJER



Page 21 of 24

Figure 7 - Top: histogram of cell by cell standard deviation of the CHM at SOAP. Bottom: histogram of

cell by cell standard deviation of the CHM at SJER



Page 22 of 24

7 VALIDATION AND VERIFICATION

To validate the CHM product ground truth measurements are required in the form of measured tree

heights. NEON will measure tree heights at selected plots for the cross-section of vegetation types at

each site (RD[11]). Ground-truth tree height measurements are obtained with an inclinometer for all

trees within pre-defined plot limits. The position of the trees are obtained from compass bearings and

measured distances referenced to a plot center location established with differential GPS. At the time of

writing, only the SJER and SOAP (D17) sites had ground height data available for comparison.

Unfortunately, the accuracy of the GPS positions of the plot center locations at SOAP was poor due to

the density and height of the surrounding vegetation which blocked the GPS signals. Therefore, a direct

comparison of ground truth measurements against the CHM was not possible. Therefore, the verification

of the CHM will be performed with only data from SJER. As additional data becomes available in the

future, validation and verification of the CHM product will be expanded.

Ground truth canopy height measurements were available from 17 plots at SJER, and comprised 362

ground height measurements. Several of the measurements were obtained from dead trees or snags,

which were removed from the analysis. As the analysis proceeded, it was found that a large number of

the results showed LiDAR heights which were much larger than the measured heights. Investigation

revealed that these often existed in areas where smaller trees existed beneath the canopy of larger trees.

As the CHM provides a representation of only the top layer of the canopy, it is unable to represent these

cases. The ground truth height data also included measurements of tree crown diameter, which allowed

the validation data to be assessed for measurements of trees that fell within the canopy radius of nearby

trees, and were shorter. Any measured trees determined to be below the upper canopy layer were

removed from the analysis. The GPS measurements of the plot center locations were known to have an

uncertainty of approximately 2 -3 m. Given that the maximum point for any given tree was measured on

the ground, ground measured heights were compared against the maximum height from the CHM within

a 3 m radius surrounding each measured tree location. The comparison resulted in two outlying points,

both of which were traced to smaller trees located on the edge of the crown of a large nearby tree. It is

believed that either the crown diameter of the taller nearby tree wasn’t accurately measured, preventing

earlier removal based on the crown diameter, or the 3 m search radius instituted due to the GPS

uncertainty included part of the taller tree crown. As a result, the two outlying points were removed for

generation of final results.

Theoretically, the relationship between the ground measured heights should produce a linear regression

with slope equal to 1 and intercept equal to zero since the same quantities (tree height) are being

measured by different instruments. After the appropriate points had been removed as explained above,

46 remained for determining a least- squares linear regression. The resulting regression produced a line

with a slope of 1.013 and an intercept of -0.493 (Figure 8), with no obvious trend evident in the residuals



Page 23 of 24

(Figure 9), and an RMSE of 0.69 m. The high correspondence between the slope of the least-squares

regression and the theoretical ideal model (1.013 vs. 1) indicates that the CHM model derived from the

LiDAR has a high relative accuracy. The 95% confidence interval of the slope parameter is ±0.045,

indicating that the theoretical value of 1 is also within the confidence limit. The intercept of the

regression is slightly low at -0.493 indicting that the LiDAR derived canopy height model is below the ideal

value of zero, and the CHM is generally underestimated. The observed underestimation is consistent with

existing reported validation efforts of LiDAR derived tree height (discussed in Section 6). The

correspondence of the NEON results with previous results indicates the CHMs produced are reasonable

approximations given the technological constraints of the airborne LiDAR sensor and processing

techniques.

Figure 8 - Linear regression of ground measured canopy heights vs. the LiDAR derived CHM at SJER



Page 24 of 24

Figure 9 - Vertical residuals of ground measured tree height vs. LiDAR derived CHM linear regression

model at SJER

8 FUTURE PLANS AND MODIFICATIONS

Currently, the primary focus of future modifications will be on improving the selection of the rasterization

threshold during the canopy height model creation. As discussed in Section 4.2, this is currently set to 3

m, and was based on the current average point density and testing on a limited number of sites. Testing

will be on-going to assess if it is possible to dynamically select this value to minimize interpolation over

legitimate canopy gaps while also adequately filling unwanted data pits. Additionally, the CHM product is

currently distributed in only geoTIFF format and future versions may also be provided in a NEON-specific

HDF5 (Hierarchical Data Format v5) format. In addition to the changes to the algorithm and product

format, validation and verification will be on-going as field data from the NEON sites become available.

Once results from additional validation and verification studies are available, changes in processing will

be undertaken if it is revealed there are opportunities to improve accuracy.

9 BIBLIOGRAPHY

Andersen, H.-E., Reutebuch, S. E., & McGaughey, R. J. (2006). A rigorous assessment of tree height

measurements obtained using airborne lidar and conventional field methods. Canadian Journal

of Remote Sensing, 32(5), 355–366.

Asner, G. P., Hughes, R. F., Mascaro, J., Uowolo, A. L., Knapp, D. E., Jacobson, J., … Clark, J. K. (2011). High-

resolution carbon mapping on the million-hectare island of hawaii. Frontiers in Ecology and the

Environment, 9(8), 434–439.



Page 25 of 24

Asner, G. P., Powell, G. V., Mascaro, J., Knapp, D. E., Clark, J. K., Jacobson, J., …, Victoria, E., et al. (2010).

High-resolution forest carbon stocks and emissions in the amazon. Proceedings of the National

Academy of Sciences, 107(38), 16738–16742.

ASPRS. (2009). Las specification, version 1.3–r10. The American Society for Photogrammetry & Remote

Sensing (ASPRS).

Ben-Arie, J. R., Hay, G. J., Powers, R. P., Castilla, G., & St-Onge, B. (2009). Development of a pit filling

algorithm for lidar canopy height models. Computers & Geosciences, 35(9), 1940–1949.

Carlson, T. N. & Ripley, D. A. (1997). On the relation between ndvi, fractional vegetation cover, and leaf

area index. Remote sensing of Environment, 62(3), 241–252.

Chasmer, L., Hopkinson, C., Smith, B., & Treitz, P. (2006). Examining the influence of changing laser pulse

repetition frequencies on conifer forest canopy returns. Photogrammetric Engineering & Remote

Sensing, 72(12), 1359–1367.

Chasmer, L., Hopkinson, C., & Treitz, P. (2006). Investigating laser pulse penetration through a conifer

canopy by integrating airborne and terrestrial lidar. Canadian Journal of Remote Sensing, 32(2),

116–125.

Gaveau, D. L. & Hill, R. A. (2003). Quantifying canopy height underestimation by laser pulse penetration

in small-footprint airborne laser scanning data. Canadian Journal of Remote Sensing, 29(5), 650–

657.

Goetz, S. & Dubayah, R. (2011). Advances in remote sensing technology and implications for measuring

and monitoring forest carbon stocks and change. Carbon Management, 2(3), 231–244.

Goodwin, N. R., Coops, N. C., & Culvenor, D. S. (2006). Assessment of forest structure with airborne lidar

and the effects of platform altitude. Remote Sensing of Environment, 103(2), 140–152.

Goulden, T. & Hopkinson, C. (2010). The forward propagation of integrated system component errors

within airborne lidar data. Photogrammetric Engineering & Remote Sensing, 76(5), 589–601.

Hopkinson, C. (2007). The influence of flying altitude, beam divergence, and pulse repetition frequency

on laser pulse return intensity and canopy frequency distribution. Canadian Journal of Remote

Sensing, 33(4), 312–324.

Hyyppä, J., Hyyppä, H., Yu, X., Kaartinen, H., Kukko, A., & Holopainen, M. (2009). Forest inventory using

small-footprint airborne lidar. Topographic laser ranging and scanning: Principles and processing,

335–370.

Isenburg, M. (2014). Rasterizing perfect canopy height models form lidar. Retrieved November 4, 2014,

from http://www.ngs.noaa.gov/GEOID/GEOID12A/

Khosravipour, A., Skidmore, A. K., Isenburg, M., Wang, T., & Hussin, Y. A. (2013). Development of an

algorithm to generate a lidar pit-free canopy height model. In Silvilaser international conference

on lidar applications for assessing forest ecosystems, beijing, china (Vol. 6, pp. 125–128).

Khosravipour, A., Skidmore, A. K., Isenburg, M., Wang, T., & Hussin, Y. A. (2014). Generating pit-free

canopy height models from airborne lidar. Photogrammetric Engineering & Remote Sensing,

80(9), 863–872.



Page 26 of 24

Lefsky, M. A., Cohen, W. B., Harding, D. J., Parker, G. G., Acker, S. A., & Gower, S. T. (2002). Lidar remote

sensing of above-ground biomass in three biomes. Global ecology and biogeography, 11(5), 393–

399.

Lefsky, M. A., Cohen, W. B., Parker, G. G., & Harding, D. J. (2002). Lidar remote sensing for ecosystem

studies lidar, an emerging remote sensing technology that directly measures the three-

dimensional distribution of plant canopies, can accurately estimate vegetation structural

attributes and should be of particular interest to forest, landscape, and global ecologists.

BioScience, 52(1), 19–30.

Liu, H. & Dong, P. (2014). A new method for generating canopy height models from discrete-return lidar

point clouds. Remote Sensing Letters, 5(6), 575–582.

Maltamo, M., Mustonen, K., Hyyppä, J., Pitkänen, J., & Yu, X. (2004). The accuracy of estimating individual

tree variables with airborne laser scanning in a boreal nature reserve. Canadian Journal of Forest

Research, 34(9), 1791–1801.

Næsset, E. (2004). Effects of different flying altitudes on biophysical stand properties estimated from

canopy height and density measured with a small-footprint airborne scanning laser. Remote

Sensing of Environment, 91(2),243–255.

Næsset, E. (2009). Effects of different sensors, flying altitudes, and pulse repetition frequencies on forest

canopy metrics and biophysical stand properties derived from small-footprint airborne laser data.

Remote Sensing of Environment, 113(1), 148–159.

Næsset, E., Gobakken, T., Holmgren, J., Hyyppä, H., Hyyppä, J., Maltamo, M., … Söderman, U. (2004). Laser

scanning of forest resources: the nordic experience. Scandinavian Journal of Forest Research,

19(6), 482–499.

Ritter, N., Ruth, M., Grissom, B. B., Galang, G., Haller, J., Stephenson, G., …, Stickley, J., et al. (2000). Geotiff

format specification geotiff revision 1.0. URL: http://www. remotesensing.

org/geotiff/spec/geotiĭome. html.

Rosenqvist, Å., Milne, A., Lucas, R., Imhoff, M., & Dobson, C. (2003). A review of remote sensing technology

in support of the kyoto protocol. Environmental Science & Policy, 6(5), 441–455.

Shamsoddini, A., Turner, R., & Trinder, J. (2013). Improving lidar-based forest structure mapping with

crown-level pit removal. Journal of Spatial Science, 58(1), 29–51.

Suárez, J. C., Ontiveros, C., Smith, S., & Snape, S. (2005). Use of airborne lidar and aerial photography in

the estimation of individual tree heights in forestry. Computers & Geosciences, 31(2), 253–262.

Turner, D. P., Cohen, W. B., Kennedy, R. E., Fassnacht, K. S., & Briggs, J. M. (1999). Relationships between

leaf area index and landsat tm spectral vegetation indices across three temperate zone sites.

Remote sensing of environment,70(1), 52–68.

Yu, X., Hyyppä, J., Hyyppä, H., & Maltamo, M. (2004). Effects of flight altitude on tree height estimation

using airborne laser scanning. Proceedings of the Laser Scanners for Forest and Landscape

Assessment–Instruments, Processing Methods and Applications, 02–06.

Zhao, D., Pang, Y., Li, Z., & Sun, G. (2013). Filling invalid values in a lidar-derived canopy height model with

morphological crown control. International journal of remote sensing, 34(13), 4636–4654.